Mnemosyne OS

What Mnemosyne OS does

Mnemosyne OS is a self-hosted memory system for AI agents that treats memory as a lifecycle of capture, distill, age, forget and resurface, rather than a plain vector store or RAG pipeline. It runs on PostgreSQL 16 with pgvector and a FastAPI service, and it plugs into Hermes as a memory provider with 14 tools, including palace_summon, plus automatic fact extraction.

Memories are organized like a palace: a taxonomy of 7 wings and 20 rooms, an archive number for each memory such as K·NET·PROXY·2026-0007, and tome cards that hold a description and a content pointer. Retrieval uses a three-channel summon called name, guide and resonate, where the name channel is an exact hit on archive number, title or tag. The stack also provides 1024-dimension HNSW vector search and BM25 plus ILIKE full-text search. Raw conversations stay lossless in Hermes state.db and are synced into PostgreSQL, and retention tiers separate permanent, long and short memories. An edge-cloud sync links SQLite and PostgreSQL.

Key features

• Hermes memory provider with 14 tools, including palace_summon
• Palace taxonomy of 7 wings and 20 rooms with archive numbers and tome cards
• Three-channel summon: name, guide and resonate
• LLM-driven fact extraction that runs after conversations
• Retention tiers for permanent, long and short memories
• Edge-cloud sync between SQLite and PostgreSQL

When to use it

• Give a Hermes agent long-term memory that survives context resets over weeks
• Keep conversation history lossless while still extracting searchable facts
• Run a private memory service for agents on your own cloud server

Who it is for: Hermes Agent users who want a structured, self-hosted long-term memory service rather than a flat vector store.

How it fits with Hermes Agent

Described as Hermes-native: it is a Hermes memory provider that reads Hermes conversation state and exposes 14 agent tools.

FAQ